基于协同网络与度量学习的标签噪声鲁棒联邦学习方法

Collaborative Network and Metric Learning Based Label Noise Robust Federated Learning Method

针对联邦学习中标签噪声问题的研究较少,目前的主流方法是,服务器端引入基准数据集对客户端的模型进行评估,对客户端的聚合权重、特征类中心进行控制等,但大多数方法区分噪声客户端/噪声样本的能力尚有提升空间。文中提出了一种基于协同网络与度量学习的标签噪声鲁棒联邦学习方法。该方法由以下3部分组成:1)客户端互评分机制:客户端为彼此模型评分,构建评分矩阵,进一步将其转化为邻接矩阵,以区分干净/噪声客户端。2)协同网络模块:通过构建两个协同对等的联邦网络模型,使用简森-香农散度为协同网络彼此的训练区分干净样本与噪声样本。3)联邦-协同网络三元组损失:为噪声样本设计损失函数,约束同一噪声样本协同网络的输出特征。在CIFAR-10和CIFAR-100两个公开数据集上进行实验验证,结果表明所提方法在准确性上具有优势。

Currently,there is limited research on the problem of label noise in federated learning.The main approaches involve introducing a benchmark dataset on the server side to evaluate the client’s model,controlling the aggregation weights and feature class centers of the clients.However,most methods still have room for improvement in distinguishing noisy clients or noisy samples.This paper proposes a label-noise robust federated learning method based on co-networks and metric learning.The method consists of the following three parts:1)Client mutual evaluation mechanism.Clients score each other’s models,construct a rating matrix,and further transform it into an adjacency matrix to differentiate clean/noisy clients.2)Collaborative network module.By constructing two collaborative equivalent federated network models,the Jensen-Shannon divergence is used to distinguish clean samples from noisy samples for the training of collaborative networks.3)Federated-collaborative network triplet loss.A loss function is designed to constrain the output features of the collaborative networks for the same noisy samples.Experimental verification is conducted on the publicly available datasets CIFAR-10 and CIFAR-100,and the results demonstrate the superiority of the proposed method in accuracy.